Enhanced Physics-Informed Neural Networks for Deep Tunnel Seepage Field Prediction: A Bayesian Optimization Approach

Yiheng Pan, Yongqi Zhang, Qiyuan Lu, Peng Xia, Jiarui Qi, Qiqi LuoWater 2025 · DOI

TL;DR

This paper introduces an enhanced Physics-Informed Neural Network (PINN) framework to accurately predict groundwater seepage around deep tunnels, a critical but challenging problem. By integrating novel techniques like adaptive sampling and Bayesian optimization, the method achieves exceptional accuracy even with limited data, offering a more robust and practical solution for geotechnical engineering.

What Problem Does This Paper Address?

Predicting groundwater seepage around deep tunnels is crucial for ensuring construction safety, long-term structural stability, and minimizing environmental impacts. Traditional methods like analytical solutions and numerical simulations (FEM/FDM) struggle with complex geometries, mesh generation, and computational limitations. Standard Physics-Informed Neural Networks (PINNs), while promising for solving partial differential equations (PDEs), face their own hurdles in this domain: training instability, difficulty in accurately handling complex boundaries (like circular tunnels), and inefficient data sampling. This paper aims to overcome these limitations by developing a more sophisticated PINN approach tailored for deep tunnel seepage prediction.

Methods

The core of this work is an "enhanced PINN framework" built upon several innovative components:

  • Geometry-Aware Trial Functions (ADF): To precisely handle circular tunnel boundaries and enforce Dirichlet boundary conditions (constant head), the authors employ a geometry-aware trial function method based on an approximate distance function (ADF). This hard-constraints the boundary conditions, avoiding the accuracy compromises of traditional soft constraints.
  • Three-Stage Training Strategy:
    • Stage 1: Adam Pre-training with Exponential Moving Average (EMA): This stage uses Adam optimization with EMA to stabilize training and smooth network weights, preventing oscillations and establishing a stable foundation.
  • Stage 2: Adam Reinforcement Training with Residual Adaptive Refinement with Distribution (RAR-D): This novel adaptive sampling technique dynamically identifies regions with high PDE residuals (where the physics is poorly satisfied) and increases the density of training points in those areas. This ensures better resolution in critical zones without uniform, computationally expensive sampling.
  • Stage 3: L-BFGS-B High-Precision Convergence Fine-tuning: A standard but effective optimization algorithm for achieving high-accuracy convergence in the final stages.
  • Bayesian Optimization (BO): This global optimization method is used to automatically tune hyperparameters, particularly the weights of different loss function components (PDE, Dirichlet, Neumann, and data). This automates a tedious manual process and helps find optimal configurations.

Datasets

The study uses a simulated dataset based on a two-dimensional steady-state groundwater flow model around a horizontal tunnel. The analytical solution for this model, derived by Lei [9], serves as the ground truth for evaluating the PINN's performance. The training data consists of:

  • Collocation Points: 5000 uniformly distributed random points within the computational domain (excluding the tunnel interior).
  • Boundary Condition Points: Points systematically generated to enforce Dirichlet (surface and tunnel boundaries) and Neumann (impermeable boundaries) conditions.
  • Monitoring Points: Randomly generated points representing actual measurement data, with their density varied for robustness analysis.

No new benchmark datasets were introduced; rather, a well-established analytical solution for a specific tunnel seepage scenario was used.

Key Insights and Ideas

  • Loss Weighting is King: Bayesian optimization revealed a critical insight: the configuration of loss function weights (how much importance is given to satisfying the PDE, boundary conditions, and data) has a much greater impact on model performance than the neural network's architecture (number of layers or neurons) in physics-constrained problems.
  • Adaptive Sampling is Crucial for Accuracy: The RAR-D method significantly boosts accuracy by intelligently focusing computational effort on regions where the physics is least satisfied, rather than relying on uniform sampling. This is particularly effective near the tunnel where gradients are steep.
  • Hard Constraints for Boundaries Pay Off: Using geometry-aware trial functions to hard-constraint the tunnel and surface boundaries dramatically improves accuracy and stability, especially compared to soft constraints.
  • Data Efficiency is Remarkable: The enhanced PINN framework demonstrates excellent performance even with very few monitoring points (as few as five), highlighting its practical applicability in data-scarce scenarios common in real-world engineering.

Key Results

  • Exceptional Accuracy: The enhanced PINN achieved an overall average relative error of 5 × 10⁻⁵, with a maximum relative error at tunnel boundaries of 6 × 10⁻³.
  • Robustness in Data Scarcity: Even with only 10 monitoring points, the model's accuracy was comparable to using 50 points, demonstrating high data efficiency. Performance degraded significantly only with five points.
  • Component Contributions: Ablation studies showed that adaptive sampling (RAR-D) provided the most significant accuracy improvement (52% error reduction), while EMA ensured training stability, and Bayesian optimization delivered the final precision enhancement.
  • Computational Cost: The full framework took approximately 3x the training time of a baseline PINN, but the accuracy gains were deemed justifiable for engineering applications.

Limitations and Future Work

  • Dimensionality and Complexity: Current validation is primarily for 2D steady-state problems under ideal homogeneous conditions. Extending to complex 3D geometries and transient scenarios remains a challenge.
  • Geotechnical Parameter Sensitivity: While Bayesian optimization explored loss function weights, it did not systematically analyze the sensitivity to geotechnical parameters like varying permeability or tunnel geometries. This is a critical gap for practical deployment.
  • Field Data Validation: The study lacks validation with real-world field data, which is often constrained by spatio-temporal resolution, measurement accuracy, geological uncertainties, and commercial confidentiality.
  • Future Directions: The authors suggest extending the framework to handle heterogeneous media, anisotropic flow, and transient problems. They also propose future research on geotechnical parameter sensitivity studies, transfer learning, and systematic field validation.

Summary updated

Abstract

Predicting tunnel seepage field is a critical challenge in the construction of underground engineering projects. While traditional analytical solutions and numerical methods struggle with complex geometric boundaries, standard Physics-Informed Neural Networks (PINNs) encounter additional challenges in tunnel seepage problems, including training instability, boundary handling difficulties, and low sampling efficiency. This paper develops an enhanced PINN framework designed specifically for predicting tunnel seepage field by integrating Exponential Moving Average (EMA) weight stabilization, Residual Adaptive Refinement with Distribution (RAR-D) sampling, and Bayesian optimization for collaborative training. The framework introduces a trial function method based on an approximate distance function (ADF) to address the precise handling of circular tunnel boundaries. The results demonstrate that the enhanced PINN framework achieves an exceptional prediction accuracy with an overall average relative error of 5 × 10−5. More importantly, the method demonstrates excellent practical applicability in data-scarce scenarios, maintaining acceptable prediction accuracy even when monitoring points are severely limited. Bayesian optimization reveals the critical insight that loss weight configuration is more important than network architecture in physics-constrained problems. This study is a systematic application of PINNs to prediction of tunnel seepage field and holds significant value for tunnel construction monitoring and risk assessment.
BibTeX
@inproceedings{pan2025enhanced,
  author = {Yiheng Pan and Yongqi Zhang and Qiyuan Lu and Peng Xia and Jiarui Qi and Qiqi Luo},
  title = {{Enhanced Physics-Informed Neural Networks for Deep Tunnel Seepage Field Prediction: A Bayesian Optimization Approach}},
  booktitle = {Water},
  year = {2025},
  doi = {10.3390/w17172621},
}
Figure